Quanta: Open-Source Python Library Unifies Quantized Vectors, BM25 and Graph RAG
_reachsumit · x · 2026-09-17
An arXiv paper presents Quanta, a self-contained open-source Python library that unifies 4-bit quantized dense vector search, BM25 full-text retrieval, and knowledge-graph traversal behind one retrieval API — replacing the typical multi-system RAG stack (ANN index + search engine + graph DB + document store).
Key design choices:
- Weighted reciprocal rank fusion instead of normalizing heterogeneous scores onto a shared range, which the author argues is ill-posed because such normalizations are query-dependent.
- Graph as candidate expander, not relevance scorer: traversal widens the candidate pool; newly admitted documents are re-scored by the dense indexes under an identifier allowlist, so structure decides what is considered while content evidence decides ranking.
More from coding & agent
- The hard part of AI coding: knowing which old best practices are now antipatterns — DavidWells · 2026-11-23
- New Developer Superpower: Learn Architecture Patterns and Steer the AI — DavidWells · 2026-11-23
- orgchat: open-source E2E-encrypted MCP relay lets Claude Code, Codex and Cursor agents talk to each other — Nice-Enthusiasm8191 · 2026-11-23
- Long Wind: a browser wuxia action game fully coded by Claude Opus 5.5 in one day — Promptmethus · 2026-11-23
- Dev ships arcade football game with Claude Code: 100 plan files, Sonnet builders, Opus reviewers — danielknugs · 2026-11-23
- Sentry CEO asks: how do you write integration tests for agents? — zeeg · 2026-11-23